基于地标特征和元学习方法推荐最适用优化算法

设计并实证研究一种基于地标特征和元学习方法推荐最佳优化算法的实现框架.地标特征摒弃了传统的问题简单特征、统计特征和信息理论特征复杂的提取过程,通过简化运行算法并仅以算法的相对性能表现作为问题特征集.在此基础上,利用元学习方法训练建模并针对新问题作出算法推荐.为验证推荐效果,以多模式资源约束的项目调度问题(MRCPSP)为优化对象,以人工蜂群、蚁群、粒子群和禁忌搜索4种元启发式算法作为推荐对象,分别使用人工神经网络、k最近邻、决策树以及随机森林4种元学习方法建立推荐元模型.计算结果表明,多种元学习方法均指向相近的推荐准确率,平均稳定在70%以上,最高可达95%.基于地标特征和元学习方法实现优化算法推荐是一个值得进一步探讨的新方向.

Recommending best suitable metaheuristic based on landmarking feature and meta-learning approach

This paper presents and empirically studies an implementation framework for recommending the best suitable optimization algorithm based on landmarking features and meta-learning approaches. The landmarking abandons the traditional feature extraction techniques and/or approach. The landmarking features are obtained using the simplified algorithm on the problem and using only the relative performance of the algorithm as the feature dataset. On this basis, meta-learning approaches are applied to train the metamodel and make algorithm recommendations for new problems. In order to verify the effect, a set of multi-mode resource constrained project scheduling problem(MRCPSP) is selected as the objective. Four meta-heuristic algorithms, namely artificial bee colony, ant colony system, particle swarm optimization and tabu search, are selected as the recommended algorithms. The four meta-learning approaches, namely artificial neural network, k-nearest neighbourhood, decision tree and random forest, are used to generate the recommended meta-model. The empirical study shows that all the prediction results point to similar recommendation accuracy, with an average stabilised around 70% and a maximum at 95%. The optimization algorithm recommendation based on the landmarking and meta-learning approach is a new direction worthy of further exploration.